Storage Class Memory: A Low Power Storage Opportunity

Transcription

1 IBM Almaden Research Center Storage Class Memory: A Low Power Storage Opportunity Rich Freitas Motivation Trends Demand for storage continues to be robust Storage performance gain has not kept pace with that of servers so proportionally more disks will be needed in the data center The amount of power consumed in data centers is becoming an issue New storage technologies: Flash, PCM, Questions Where will disk storage be in 2020? Will new storage technologies help? 2 1

2 Agenda Technology and power consumption Disks Flash Phase Change Memory Illustrative power comparison High data-rate and high transaction-rate scenarios Power comparisons for technologies 3 HDDs Actuator Arm Magnetic head (slider) Invented in the 1950s Mechanical device consisting of a rotating magnetic media disk and actuator arm w/ magnetic head HUGE COST ADVANTAGES $ High growth in disk areal density has driven the HDD success $ Magnetic thin-film head wafers have very few critical elements per chip (vs. billions of transistors per semiconductor chip) $ Thin-film head (GMR-head) has only one critical feature size controlled by optical lithography (determining track width) $ Areal density is control by track width times (X) linear density 4 2

3 HDD strong suit is Areal Density 5 Future of HDD Higher densities through perpendicular recording Wikipedia Jul Gb/in 2 ~4 TB patterned media pm images/conventional_pattern_media.pdf 6 3

4 But, problems with Disk Drive Maximum Sustained Data Rate Bandwidth [MB/s] MB/s Date Date Available 7 And Disk Drive Latency Latency [ms] Date 8 4

5 So, Bandwidth Problem is getting much harder to hide with parallelism Access Time Problem is also not improving with caching tricks Power/Space/Performance Cost 9 Typical Storage Device Power Watts Drive type size stand-by idle typical start-up 3.5" 15K RPM FC/SAS 300 GB " 10K RPM FC/SAS 300 GB " 10K RPM FC 300 GB " 7200 RPM SATA 500 GB " 7200 RPM SATA NL 500 GB " 7200 RPM SATA 500 GB " 15K RPM SAS 2.5" 10K RPM SAS 73 GB 73 GB 30% less than 15K 3.5 SAS " Mobile 7200 RPM SATA 100 GB " Mobile 5400 RPM SATA 100 GB USB Flash Disk 32 GB " laptop SSD 73 GB " Enterprise SSD 155 GB

6 Illustrative Disk Drive Power Profile Standby power Both spindle and actuator Motors off Base electronics powered Idle Spindle motor running Actuator motor off Most electronics powered Typical Spindle motor running Actuator motor in use periodically All electronics powered Startup spindle motor starting ~30 seconds for startup Typ pwr 0 idle power: always being consumed regardless of whether operations are in progress or not 11 Many Competing Technologies for SCM Phase Change RAM most promising now (scaling) Magnetic RAM used today, but poor scaling and a space hog Magnetic Racetrack basic research, but very promising long term Ferroelectric RAM used today, but poor scalability Solid Electrolyte and resistive RAM (Memristor) early development, maybe? Organic, nano particle and polymeric RAM many different devices in this class, unlikely Improved FLASH still slow and poor write endurance bistable material plus on-off switch Generic SCM Array 12 6

7 What is Flash? gate oxide Based on MOS transistor source drain Transistor gate is redesigned Floating Gate source control gate e - e - Flash Memory 1 drain Charge is placed or removed near the gate The threshold voltage V th of the transistor is shifted by the presence of this charge Floating Gate source control gate e - e - e - e - e - drain Flash Memory 0 The threshold Voltage shift detection enables non-volatile memory function. Single Level vs Multi-level cell Scaling issues 13 Feeds and Speeds for typical NAND Flash Cell Size Read Access Time Read Write Erase Start Up Time Market Size (2007) Applications NAND 4 F 2 (2 F 2 virtual x 2-bit MLC) us MB/s 5-8MB/sec 2ms us $14.2B Multimedia 14 7

8 Representative NAND Flash Device Interface: one or two bytes wide Transition to ONFI for some vendors Data accessed in pages 2112, 4224 or 8448 Bytes Data erased in blocks Block = s Power circuits Charge Pumps Clock drivers Etc. B U F C BA C UH E F Power Circuits Block 0 Block 1 Block 63 ONFI Open NAND Flash Interface 15 Illustrative Flash Read Power Profile 20us 100us 1 A 2 A 3 A 4 480us 8192B 30 mw idle idle Read Access 2 KB Data transfer 16 8

9 Illustrative Flash Write Power Profile 1200us 100us 200us ,000us 30 mw 1 mw idle idle 2 KB Data transfer Erase block Program page 17 Typical Storage Device Power Watts Drive type size stand-by idle typical start-up 3.5" 15K RPM FC/SAS 300 GB " 10K RPM FC/SAS 300 GB " 10K RPM FC 300 GB " 7200 RPM SATA 500 GB " 7200 RPM SATA NL 500 GB " 7200 RPM SATA 500 GB " 15K RPM SAS 2.5" 10K RPM SAS 73 GB 73 GB 30% less than 15K 3.5 SAS " Mobile 7200 RPM SATA 100 GB " Mobile 5400 RPM SATA 100 GB USB Flash Disk 32 GB " laptop SSD 73 GB " Enterprise SSD 155 GB

10 Illustrative Flash SSD Design CPU MEMORY 1 DRAM Flash Control Flash Flash Flash 2 Flash Flash Flash n Idle power: 5 W Flash Flash Flash Typical operating power: 8 W Many Competing Technologies for SCM Phase Change RAM most promising now (scaling) Magnetic RAM used today, but poor scaling and a space hog Magnetic Racetrack basic research, but very promising long term Ferroelectric RAM used today, but poor scalability Solid Electrolyte and resistive RAM (Memristor) early development, maybe? Organic, nano particle and polymeric RAM many different devices in this class, unlikely Improved FLASH still slow and poor write endurance bistable material plus on-off switch Generic SCM Array 20 10

11 Phase-Change RAM Bit-line PCRAM programmable resistor Word -line Access device (transistor, diode) Voltage temperature RESET pulse T melt Potential headache: High power/current affects scaling! SET pulse T cryst time Potential headache: If crystallization is slow affects performance! 21 PCM performance and power SWAG Performance 15 us 1 us 1 us For devices optimized for storage applications: RD Read access time of 1 us Write access time of 3-5 us Interface LPDDR or similar Power 32 us Transfer power 15mW 16 us Read power 20 mw write and transfer write write Write: 85 mw Write-in-place S0, no erase 22 11

12 Illustrative PCM SSD Design CPU MEMORY SWAG Rd: 2400 MB/s 0.6W 600,000 IOPS Wt: 840 MB/s 2.0 W 210,000 IOPS 1 SCM SCM Control SCM SCM SCM SCM SCM SCM 2 n SCM SCM SCM Idle power: 1 W Typical operating power: 3 W 23 Device comparison: energy and power IOPS BW Idle power Typ. power Energy per Rd op in 1 sec 10% utilization 50% 90% Energy is what you pay for In 2020 a disk will cost ~3000x more per operation than a PCM SSD IOPS, cost, power power, cost IOPS Today K 485,000 uj 125,000 uj 85,000 uj 105,000 uj 25,000 uj 16,000 uj 1,060 uj MB/s 9 W 16 W disk K MB/s 4 W 6 W Flash SSD today , MB/s 5 W 8W 260 uj 171 uj PCM SSD projected , MB/s 1 W 3 W 20 uj Flash SSDs may not exist in 2020 Power for controllers, etc. viewed as second order effects 7 uj 5 uj 12

13 Extrapolate Storage Performance to 2020 Start with ASCI Purple class machine Trend for storage performance growth: 70% CAGR Driven by application requirements and investment Could be impacted by changing in architecture Assume 50 yr disk trends continue and no game changing technology invention Semiconductor technologies stay on a roughly 40% CAGR PCM edges out FLASH for the solid state storage crown Bandwidth: 0.4 TB/s 400 TB/s Transaction rate: 2 MIOPS 2000 MIOPS 25 The Promise of Solid State Disk By 2020, Storage Class Memory should revolutionize data centers Bandwidth Driven Storage System: 400 TB/s DISKS Floor Space Power 6,000 KW 6000 Square Feet 26 13

14 The Promise of Solid State Disk By 2020, Storage Class Memory should revolutionize data centers Bandwidth Driven Storage System: 400 TB/s SCM Floor Space Power 42 KW 85 Square Feet 27 The Promise of Solid State Disk By 2020, Storage Class Memory should revolutionize data centers Transaction Rate Driven Storage System: 2000 MOP/s DISKS Floor Space Power 22,000 KW 23,000 Square Feet 28 14

15 The Promise of Solid State Disk By 2020, Storage Class Memory should revolutionize data centers Transaction Rate Driven Storage System: 2000 MOP/s SCM Floor Space Power 1 KW 11 Square Feet 29 My Conclusions Idle power for disks drives up their energy costs significantly Energy per op for SSDs is much better than for HDDs SSDs gain because of their very high transaction rates and and high bandwidth Combined with generally lower idle power But, even so, SSDs have a higher idle power that one would expect for solid state technology Careful design of SSDs should alleviate much of this Current trends would seem to indicate that Enterprise disks may not be widely used in 2020 and their position will be taken by Solid State Storage 30 15

16 Questions? Break Time 31 Further reading IBM Journal of Research and Development, special issue on Storage Technologies and Systems Volume 52, Number 4/5 July/September 2008 R. F. Freitas and Winfried Wilcke, Storage-class Memory: The next storage system technology, pg G. W. Burr, etal., Overview of candidate device technologies for storage-class memory, pg S. Raoux, etal., Phase-change random access memory: a scalable technology, pg Other papers Grupp, Laura M., etal., Characterizing Flash Memory: Anomalies, Observations, and Applications, MICRO 09, December 12 16, 2009, New York, NY, USA., pg: Lee, Kwang-Jin, etal., A 90 nm 1.8 V 512 Mb Diode-Switch PRAM With 266 MB/s Read Throughput, IEEE JOURNAL OF SOLID-STATE CIRCUITS, VOL. 43, NO. 1, JANUARY 2008, pg Chen, Feng, etal., Understanding Intrinsic Characteristics and System Implications of Flash Memory based Solid State Drives, SIGMETRICS/Performance 09, June 15 19, 2009, Seattle, WA, USA. Copyright

17 Input from the device cost crystal ball $100, $10, $1, $ $10.00 DRAM $1.00 Flash SRAM MRAM $0.10 DRAM-pr FLASH-pr SCM $ Input Price form Trends: the Magnetic Subsystem disks and Price Solid State Crystal Disks Ball $ Ent Flash SSD PCM SSD Hi. Perf. Flash SSD $10.00 Ent. Disk SATA Disk Price: $/GB $1.00 $0.10 SSD Price = multiple of device cost 10x -40% CAGR (4.5F 2 ) 3x -40% CAGR (2.3 --> 1.1F 2 ) 3x -40% CAGR (1.0 F 2 ) $